[evan_au (OP)]

Some researchers are using AI techniques to analyze the sound of a forced cough, in an attempt to diagnose SARS-COV2.
- They claim to be able to diagnose asymptomatic cases too...
- In fact, some researchers claim that asymptomatic cases are easier to detect, because the symptomatic COVID-19 cases are hard to disentangle from colds, flu, hayfever or asthma!

If the claims are fulfilled in practice, this could be as accurate as some of the antibody tests that have been promoted in the past 6 months, as well as quicker (30s vs 15 minutes), cheaper (no disposable devices), no need for trained medical staff, and probably less intrusive (able to be performed at home every day before you go to work or school).

I'm not sure how traditional medical regulatory bodies would deal with a downloadable app developed amongst the empty pizza boxes and cola cans...
- Compared to a traditional medical device developed and manufactured under cleanroom conditions by people with medical PhDs...

What do you think?
See: https://spectrum.ieee.org/the-human-os/artificial-intelligence/medical-ai/ai-recognizes-covid-19-in-the-sound-of-a-cough

[Petrochemicals]

I'm sure its been done already.

www.express.co.uk/news/science/1263499/coronavirus-test-COVID-Voice-Detector-app-symptoms-carnegie-mellon/amp

[alancalverd]

There are two approaches to medical device regulation.

The European authorities require demonstration of safety (which is not a problem with a non-contact app) and manufacturing quality assurance - again not a problem. Whether it works is a minor consideration.

US FDA regulators require proof of "equivalence or better", and this is where the app may stumble. In the absence of an equivalent, you need to demonstrate efficacy, and if you are using a test for screening for an easily transmissible disease, you need a very high level of confidence in your negative scores.

There is also a problem of legal status. There being no cure for COVID, the only means of control is sequestration of carriers. Nobody is going to risk having somebody else contaminate his expensive phone, so it has to be a self-test and self-report (with no quality control - you don't want a potential carrier to force a cough in front of a nurse)  with the outcome being either "possibly not infectious" or "unclean". What next? Some might sequester themselves unnecessarily but I suspect the majority of Republicans and Conservative back-benchers will ignore the positive result in the name of Freedom and the Holy Economy or continue to expose themselves and others to unnecessary risk. There were all sorts of concerns about privacy when track-and-trace apps were being developed, so the best likely outcome would be some dodgy statistics. 

[Bored chemist]

Conservative back-benchers will ignore the positive result in the name of Freedom and the Holy Economy or continue to expose themselves

When you talk about  Tory MPs exposing themselves, do you mean like this?

https://www.theguardian.com/politics/2020/feb/25/tory-mp-james-grundy-apologises-after-exposing-himself-in-pub
And  this
https://www.mirror.co.uk/news/uk-news/disgraced-tory-brooks-newmark-caught-4421549

Just checking.

[alancalverd]

...the  very people who get exercised about privacy!

In the video, he asks for assurance that his face can’t be seen: “We’re OK, aren’t we? There’s no way you can get my face in?”

so you can't use a voice recognition app.

[alancalverd]

Serious postscript.

I understand Elon Musk took four standard COVID  tests in one day, administered by the same nurse, and received two positive and two negative results.

I propose to register a 10 cent coin as a medical device with the FDA as "99.999% reproducible device delivering equivalent performance with less skill and reduced likelihood of secondary infection". Should pass on the nod.
, 

[evan_au (OP)]

A phone app has very different issues than a traditional medical device:
- "Clinical Testing" means running the algorithm on a set of sound samples from people who are known to have or not have COVID-19.
- You could use the same cohort in the testing and training of multiple tools from different developers
- You could increase the size of your sample by capturing sound samples from people who are being tested for SARS-COV2. Upload the sound samples immediately, and annotate them within 24 hours as being either positive or negative for SARS-COV2 (as indicated by the PCR test).
- "Manufacture" means putting it in the App store, where millions of people can download it.

The fundamental rule of AI development is that you must not test the AI using the same data that you used for training, as the AI tends to learn to recognize individual samples and their result. What you want it to do is to identify general patterns which can be applied to data different from the set it was trained on.
- You normally randomly divide the input data into two groups: train on one set of data, and test on the independent set of data.